search

dasher-project / dasher-web

Dasher text entry in HTML, CSS, JavaScript, and SVG
https://dasher-project.github.io/dasher-web/browser/
MIT License
46 stars 8 forks source link

Integrate the Language Model - Google Research Repository #53

Open jcope opened 3 years ago

jcope commented 3 years ago

https://github.com/google-research/mozolm

sjjhsjjh commented 3 years ago

Maybe here in the issue, or in a design document, could the following be answered please.

willwade commented 3 years ago

So. We could run a server for the gRPC service. But there would be some major issues to check on privacy. But I think we should consider them. It might need a decent payment scheme behind it too to fund its running particularly if commercial entities use it. But the privacy. That’s a big thing.

agutkin commented 3 years ago

Maybe here in the issue, or in a design document, could the following be answered please.

  • How does Mozolm handle persistent storage?

At the moment the microservice will load whichever model and/or training (in case of PPM) files upon initialization. We haven't implemented the persistence in the microservice API, but it is there on a model implementation level - unlike the JavaScript implementation that uses a non-persistent suffix trie backend, Brian's implementation uses finite-state transducer which is trivially savable to disk. All we need is to hook-up a network request to flush the current state of the model to disk.

  • Could there be a public or other hosted service that runs the Mozolm server?

Yes.

agutkin commented 3 years ago

Cc-ing Brian (@roark-google) on this.